Proposal for Cloud Computing R&D in

ATLAS Distributed Computing

1. Introduction
The ATLAS Computing Model was designed around the concepts of grid
computing to combine the resources of more than one hundred Worldwide LHC
Computing Grid (WLCG) sites for ATLAS offline data storage, distribution,
processing and analysis. After more than a year in full operation, the system has
demonstrated that it can successfully cope with the experiments computing
requirements during data-taking.

However, a new and emerging paradigm in the delivery of IT services cloud
computing presents an improved approach to managing and provisioning
resources, thereby allowing applications to easily adapt and scale to varied usage
demands. By providing an Infrastructure as a Service (IaaS), clouds aim to
efficiently share the bare hardware resources for storage and processing without
sacrificing flexibility in services offered to applications. A new and increasingly
competitive market has emerged to offer these cost effective computing
resources, and companies small and large already make extensive use of them.

It is therefore of interest for an ATLAS Distributed Computing (ADC) R&D project
to investigate opportunities for adapting its existing services to make use of
cloud computing resources and target an upgrade of the ATLAS distributed
computing technologies during the LHC shutdown in 2013.
The present document gives an initial proposal to research cloud computing for
ADC, and then to design and implement cloud-awareness in the Distributed Data
Management (DDM) system, the Production and Distributed Analysis (PanDA)
system, and in related tools and services.
2. Goals and Scope of the Project
The primary goals of this R&D project are
to evaluate the available cloud technologies in relation to the use-cases
presented by ATLAS data management, processing and analysis,
to design a model for transparently integrating cloud computing
resources with the ADC software and service stack
to implement the ATLAS cloud computing model into DDM, PanDA and
related tools and services.
Throughout this activity, collaboration with related groups in WLCG, CERN-IT,
tier centers, other experiments and existing research clouds is desired in order
to consolidate efforts and leverage commonalities between the interested parties.
Some of the potential ADC use-cases which will be investigated include:
Monte Carlo simulation on the cloud with stage-out to traditional grid
storage or long-term storage on the cloud
Data reprocessing in the cloud (with a strong caveat related to cost; see
next paragraph)
Distributed analysis on the cloud using data which is accessed remotely to
the grid sites, or analysis of data which is located in the cloud
Resource capacity bursting which is managed centrally (e.g. to handle
urgent reprocessing tasks) or regionally (e.g. to handle urgent local
analysis requests)
It is important to note that the relative attractiveness of cloud computing for
each use-case varies when considering the resources which are provided by
research and academic institutions versus commercial entities. In the former
case, the usage of cloud APIs can be seen as an alternative to grid middleware for
providing remote access to a site while better managing its local resources; one
could say that ADC should be able to use a WLCG site that chooses to make itself
available via a cloud API. In contrast, commercial cloud computing presents a
mechanism to rapidly scale up the overall ADC computing capacity; however, the
present project must consider that all operations with commercial clouds would
be chargeable and therefore the resources must be used in a way that optimizes
cost effectiveness. Working toward a model for the cost effective use of
commercial cloud resources will be a main goal of this project.
3. Roadmap
Here we present a roadmap outline for this R&D project. The general timing of
the roadmap is that exploratory work will occur during summer 2011,
concluding with a cloud computing model document in early fall. Development
will follow. Also note that some items below (notably 1,2, and 3) can happen in
parallel.
1. Basic Research
a. Review the work already carried out within ATLAS, CERN IT, WLCG,
sites, EGI and OSG
b. If possible, organize workshop to collect information about existing
cloud computing activities in collaborating organisations
2. Implement primitive data management and job execution on the cloud
a. Virtual machines
i. CERNVM as a platform, which gives access to ATLAS tools and
software
b. Evaluate potential resources
i. Various sites (e.g. Magellan at ANL, lxcloud at CERN, BNL cloud,
other cloud infrastructures related to WLCG (e.g. in Canada),
commercial clouds, etc)
ii. Various cloud APIs (Amazon (EC2, S3, etc.), OpenStack, Nimbus,
etc). Need to understand the long-term sustainability of the
APIs.
c. Implement primitive functionalities
i. Move data in and out of the cloud
ii. Execute basic jobs on the cloud
3. Use-cases study
a. Evaluate the ADC use-cases in relation to both commercial cloud
computing. Estimate costs for various models.
b. Explore new use-cases presented by cloud computing.
4. Design of the Cloud Computing Model
a. The goal of this document will be to:
i. describe how ATLAS can make use of both pledged and
chargeable cloud resources
ii. present strategies to minimally impact the existing services so
that the usage of cloud resources would be transparent to end-
user physicists
iii. present cost-effective models for the various ADC use-cases on
commercial clouds
iv. incorporate legal and security considerations
5. Development
a. Initial ideas are detailed in the section below.
6. Testing
a. Integration of the cloud with existing monitoring services, functional
tests (DDM, analysis, production) for stability and reliability
evaluations
b. Perform stress tests of the cloud solutions to study performance
4. Development Areas
4.1. Changes in ATLAS DDM
ATLAS Distributed Data Management needs to be adapted to the cloud
computing infrastructure. This development process will include the
implementation of plugin libraries that support file transfers into and out of the
cloud and that will be used in DDM Site Services and the DQ2 clients. Since there
is not a unified cloud standard, this step might require the evaluation/usage of
different cloud APIs (Amazons S3, Open Cloud Computing Interface etc.)
depending on the chosen technologies in the available resources. This work may
create requirements for the FTS development team.
A different point to address is the bookkeeping of location information for ATLAS
datasets stored in the cloud. It has to be seen if the cloud offers reliable file
catalogues and how to integrate these with the existing DDM catalogues.
On a different note, the usage of existing commercial cloud services for content
delivery may be investigated. These services allow data located in the cloud to be
automatically replicated around the world; this may have applications in ATLAS
in relation to data distribution for analysis and other hot data use-cases.
4.2. Changes in the workload management
We envisage that some development will be necessary in the PanDA pilot and
server. The PanDA pilot will need a mover module which can stage-in (e.g. wget)
cloud resident data and stage-out and register output files. The PanDA server
should not require major developments in the early phases of this project; one
could start with a set of cloud-based PanDA sites/queues, which would be
transparent to the server. In later phases of the project we envisage
opportunities for
Cost-aware job brokerage: PanDA could consider the estimated financial
cost when selecting sites for a production task or analysis jobset.
Automatic cloud resource provisioning: Panda would request new or
larger cloud capacity subject to the global workload and cost
considerations.
Further to this last point, a tool which enables automatic resource provisioning
would need to be developed. This tool could be used by services such as PanDA
or also human ADC operators to provision resources on-demand and
automatically configure them for ADC applications.
4.3. Integration with the monitoring and information systems used in
ADC
It will be necessary to add metadata about the cloud-based resources into the
monitoring and information systems that ADC depends on (Dashboards, BDII,
AGIS, PanDA schedconfig, TiersOfATLAS). This point will depend fully on the
results of the design of the Cloud Computing Model.
5. Conclusions
This proposal presents the goals and a roadmap for the introduction of cloud
computing into ADC in collaboration with CERN IT. After an initial exploration
and prototyping phase, this project will introduce a Cloud Computing Model for
ATLAS and deliver an initial implementation of the work. Development
requirements will be placed on many areas in ADC, and collaboration with the
larger team of ADC developers will be necessary. Finally, collaboration with
other groups in the WLCG will be emphasized in order to find solutions which
can be maintained and therefore sustained by a larger community.